chilicheeselies

This is the way. Cat is out of the bag now so need to figure out responsible usage.

Fortunately we have a method for that for a long time now: PRs. If you approve it, you are responsible for that approval.

This is already a thing

They will go down kicking and screaming. A company like flock has tons of investor capital. That results in a board putting tons of pressure on executive leadership. They will never go "oops nevermind this was a bad idea". It's just a other front in the class war against the billionaire Epstein class.

All agents are pretty much the same. Think of it like pair programming with someone, except that someone doesn't have feelings and you can micromanage them.

In general, I give it instructions for what I want to accomplish. It has a "plan mode" that basically instructs the LLM to give me an execution plan to approve before actually doing it. We iterate on the implementation plan together and then when I satisfied I let it generate code.

It generated diffs essentially that I can approve or deny directly, in aggregate or by individual chunk.

I can (and do) provide custom instructions that it loads whenever I start a session. Instructions are basically md files, but it can be any text.

It's a very different way of writing code, but if you ever pair programmed with a knowledgeable junior then that's kind of what it feels like.

You are both right. An LLM inherently has access to stuff the same way a brain in a jar has access to stuff. It's information comes from fine-tuning the models to return syntax that agent code can interpret as a request to invoke a tool. That tool returns information to the context of the conversation. It doesn't learn and it can't truly remember things. Every time you start a session it is brand new. It sees your codebase for the first time every time.

The information access they have is whatever the agent allows it to access via tool exposure. Be it built in tools, or MCP servers

This is a good thing to do if your goal is to gain a deep understanding of something.

If it's just to get it done, I just enforce TDD on my agent and review it's output. I don't need to be an expert in everything (and I am very much a generalist). But if you focus on a very specific thing and only that thing, then yeah what you are doing is a great way to truly understand it. It's slow, but it's totally valid.

The same reason pair programming exists. You get a better outcome when you have someone (or these days, something) to validate and implement with.

So my preferred agent atm is GitHub Copilot as a jetbrains plugin. The other day when GitHub had a 6 hour outage and I couldn't log in, my first instinct was "guess I can't work today". I realized almost immediately how fucking dumb that initial thought was. Like I had forgotten that I've been doing this for over 20 years, and around 17 of those with no AI whatsoever.

Was I slower that day? Yes absolutely. Was I incapable? Not at all.

I wouldn't want to do without it, because on the positive side it has helped me discover libraries and tools I had no idea existed, and it's made me a better archtitect. But, maybe we should still actually touch the code a little.

If you see people wearing meta glasses, make fun of them. It should be the latest most uncool thing ever. DO NOT fuck them.

How does this survive variation in temperature? Also I wonder if it's possible to fine tune this behavior out.

If you are building a new machine for home use and not starting with Linux it's crazy to me. The worst that can happen is you try it for a month or two, hate it, and then buy a windows license.